Learning Multi-Humanoid Pickup and Transport via Decentralized Object-Centric Control
Bikram Pandit, Mohitvishnu S. Gadde, Aayam Kumar Shrestha, Alan Fern
cs.RO · cs.AI
Quick verdict
Trained with at most 2 humanoids, decMHT transfers zero-shot to 10-robot teams and real Digit hardware; strong sim results but no CIs, qualitative-only hardware demo.
Summary
The paper introduces decMHT, a decentralized, object-centric framework for cooperative humanoid pickup, transport, and robot-to-robot handover. Each robot receives an attachment spec (attachment-frame pose, hand separation, frontal clearance) and an object-centric velocity/yaw/height command transformed into per-robot local commands via fixed rigid transforms; a shared two-layer-LSTM high-level policy over a frozen locomotion policy runs at 50 Hz on local observations only, with no inter-robot communication. Training uses IPPO in IsaacLab on Digit V3 through a six-stage curriculum: five single-robot stages followed by two-robot fine-tuning; teams larger than two are never trained. In simulation, the policy picks up payloads with above 97% success for teams of 1-10 robots, with RMS tilt falling from 8.04 degrees (N=1) to 1.21 degrees (N=10), lifts masses up to 50 kg zero-shot despite training on 0.5-1.5 kg boxes, and matches centralized and object-specialized per-object baselines on five unseen geometries (worst case: 85.2% pickup on a 40 kg bed). Two-robot fine-tuning mainly improves tracking, tilt, drop rate, cost of transport, and dynamic handover success (96.5% vs 87.5%). Two real Digit V3 robots demonstrate pickup, cooperative transport, and handover on a cardboard box using external motion capture. Limitations: gripperless pinching only, externally specified attachment assignment and handover timing, external pose sensing, no statistical uncertainty reporting, and qualitative-only hardware evaluation.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single learned decentralized policy trained only with one- and two-robot episodes coordinates teams of up to ten humanoids zero-shot across team sizes, attachment layouts, and object geometries.
- Policies trained only on single-robot pickup already transfer nontrivially to cooperative settings, but two-robot fine-tuning further improves tracking quality, load sharing, cost of transport, and handover robustness.
- The attachment-spec abstraction generalizes to unseen objects and masses up to 50 kg without per-object retraining, performing comparably to centralized and object-specialized baselines retrained on each evaluation object.
- The same controllers transfer without fine-tuning to real Digit V3 hardware for single-robot pickup, cooperative two-robot transport, and robot-to-robot handover using external pose sensing.
Strengths
One interface unifies three manipulation modes
The attachment spec plus object-centric command (Sec. III) covers single-robot pickup, multi-robot transport, and handover with a single policy and no per-task redesign; Sec. VIII demonstrates all three regimes on hardware with no changes to policy, inputs, or training.
Zero-shot team-size scaling from two-robot training
Trained with at most two robots (Sec. VII), the policy achieves 97.3-99.9% pickup success for N=1..10 (Table I), tilt improves monotonically with team size (8.04 to 1.21 deg RMS), and fixed payloads up to 50 kg are lifted (Fig. 3) despite a 0.5-1.5 kg training mass range (Sec. V-A).
Baselines and ablation are well designed
Honest characterization of coordination cost
Practical, compute-modest training recipe
Weaknesses
No statistical uncertainty anywhere
Hardware validation is qualitative only
Centralized baseline dropped on the hardest case
System autonomy is narrower than the framing suggests
Heavy-mass and anomaly results are figure-only or unexplained
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is decMHT?
decMHT (decentralized multi-humanoid transport) is a learned control framework from Oregon State University researchers in which each humanoid robot runs the same local policy, guided only by an 'attachment spec' describing where to grip the shared object and an object-motion command transformed into its own local frame. Robots use no inter-robot communication, so the same policy scales across team sizes.
How many humanoid robots can cooperate with this method?
In simulation, a single policy trained with at most two robots coordinates teams of up to ten humanoids with over 97% pickup success across team sizes, and handles objects up to 50 kg with larger teams. The physical demonstrations use two Digit V3 humanoids.
Did the method work on real robots?
Yes. Two real Digit V3 humanoids performed single-robot pickup-and-carry, two-robot cooperative transport, and robot-to-robot handover of a cardboard box with no real-world fine-tuning. Caveats: object pose came from an external motion-capture sensor, and results are shown in a supplementary video rather than reported as quantitative success rates.
How does decentralized multi-humanoid training compare to a centralized controller?
On five unseen objects, the decentralized policy matched centralized and object-specialized baselines that were retrained separately for each object, achieving comparable pickup success and tracking without per-object retraining; the main exception is a 40 kg four-robot bed task (85.2% pickup) where the centralized baseline was not evaluated.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-4 appear only as axis-label fragments in the extracted text; quantitative claims drawn from them (heavy-mass curves in Fig. 3, CoT values in Fig. 4) could not be verified numerically.
- The paper repeatedly defers to an 'extended paper and code release' for full implementation details (reward weights, PPO hyperparameters); that material was not available for this review.
- The supplementary video containing all hardware evidence could not be reviewed from text; hardware conclusions rest entirely on its content.
- No appendix or hyperparameter tables were present in the extraction, so some training details may have been truncated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.